Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A method comprising: storing, by a first circuitry, a parameter vector representing an energy function of a network having a plurality of visible units connected using links to a plurality of hidden units, each link being a relationship between two units; and receiving, by the first circuitry, state vectors from a second circuitry and applying an algorithm to produce new data, wherein the state vectors produced for the network by the second circuitry are sampled from a probability distribution defined by the parameter vector produced by the first circuitry; wherein the sampling further comprises generating candidate state vectors; and wherein the sampling to generate the candidate state vectors is performed independently from an inner loop of the first circuitry used to produce the new data using the state vectors, wherein the inner loop processes the state vectors to iteratively adjust the parameter vector to converge a marginal probability distribution over the visible units towards an example probability distribution; wherein the sampling and producing the new data is performed in parallel and asynchronously.
2. A method according to claim 1 , comprising sampling and/or producing the new data until a predetermined condition is reached, at which time data is exchanged from one circuitry to the other circuitry, wherein the predetermined condition is based on the converging of the marginal probability distribution over the visible units towards the example probability distribution.
This invention relates to machine learning systems, specifically methods for training neural networks or other probabilistic models by exchanging data between two parallel processing circuits. The problem addressed is the inefficiency and instability in training deep generative models, where learning may stall or diverge due to poor initialization or gradient-based optimization challenges. The method involves two parallel processing circuits operating simultaneously. Each circuit independently samples or generates new data while monitoring the marginal probability distribution over visible units (input variables). The process continues until a predetermined condition is met: the marginal probability distribution in one circuit converges toward the example probability distribution (the target data distribution). Once this condition is satisfied, data is exchanged between the circuits, allowing each to refine its learning from the other's progress. This exchange helps avoid local optima and accelerates convergence by leveraging complementary information from both circuits. The approach is particularly useful in training deep belief networks, Boltzmann machines, or other models where traditional gradient-based methods struggle. By dynamically exchanging data based on probabilistic convergence, the method improves training stability and efficiency compared to conventional techniques. The invention ensures that learning progresses toward the true data distribution by leveraging probabilistic criteria rather than fixed iteration counts or heuristic thresholds.
3. A method according to claim 2 , wherein applying the algorithm comprises applying a learning algorithm to produce an updated parameter vector that is subsequently sent to the second circuitry for re-sampling when the predetermined condition is reached.
This invention relates to adaptive signal processing systems, particularly for optimizing signal sampling in real-time applications. The problem addressed is the need for dynamic adjustment of sampling parameters to improve signal quality or efficiency, especially in systems where conditions change over time. The method involves a system with at least two processing units. The first unit applies a learning algorithm to analyze sampled data and generate an updated parameter vector. This vector is used to adjust sampling parameters in the second unit, which performs the actual signal sampling. The learning algorithm updates the parameters based on a predetermined condition, such as a threshold error rate or a time interval. When this condition is met, the updated parameters are sent to the second unit to refine the sampling process. The learning algorithm may use techniques like machine learning or statistical optimization to determine the optimal parameters. The second unit then re-samples the signal using these updated parameters, improving accuracy or reducing computational overhead. This adaptive approach ensures the system remains efficient and responsive to changing conditions without manual intervention. The invention is particularly useful in applications like sensor networks, communication systems, or industrial monitoring, where real-time adaptation is critical. By dynamically adjusting sampling parameters, the system can maintain high performance under varying environmental or operational conditions.
4. A method according to claim 3 , comprising sending the updated parameter vector to the second circuitry for re-sampling when a predetermined plural number of iterations of the learning algorithm have been performed.
This invention relates to machine learning systems, specifically methods for optimizing parameter vectors in iterative learning algorithms. The problem addressed is the computational inefficiency and potential inaccuracies that arise when parameter vectors are not properly updated or re-sampled during training. The invention provides a solution by dynamically adjusting the parameter vector based on the number of iterations performed by the learning algorithm. The method involves a system with first circuitry that executes a learning algorithm to generate an updated parameter vector. This circuitry is connected to second circuitry responsible for re-sampling the parameter vector. The key improvement is that the updated parameter vector is only sent to the second circuitry for re-sampling after a predetermined number of iterations of the learning algorithm have been completed. This ensures that re-sampling occurs at optimal intervals, balancing computational efficiency with accuracy. The predetermined number of iterations can be set based on factors such as convergence rates, computational constraints, or desired precision levels. This approach prevents unnecessary re-sampling, reducing computational overhead while maintaining model performance. The method is particularly useful in large-scale machine learning applications where iterative training is resource-intensive.
5. A method according to claim 1 , wherein the algorithm takes as input the state vectors and the parameter vector to generate output data representing a probability distribution over all possible states of the visible units.
This invention relates to machine learning systems, specifically probabilistic models for generating output data representing state probabilities of visible units in a network. The problem addressed is efficiently computing probability distributions over possible states of visible units in a system where the state is influenced by both visible and hidden units, with the goal of improving accuracy and computational efficiency in generative models. The method involves an algorithm that processes state vectors and a parameter vector to produce output data representing a probability distribution over all possible states of the visible units. The state vectors describe the current state of the system, while the parameter vector contains learned or predefined parameters that influence the probability distribution. The algorithm computes these probabilities by evaluating the contributions of both visible and hidden units, where hidden units are latent variables that are not directly observed but influence the visible units. The method ensures that the generated probability distribution accurately reflects the system's dynamics while optimizing computational resources. This approach is particularly useful in generative models, such as restricted Boltzmann machines or deep belief networks, where understanding the probability distribution of visible units is critical for tasks like data generation, feature extraction, or anomaly detection. The invention improves upon prior methods by providing a more efficient and accurate way to compute these distributions, enhancing the overall performance of probabilistic models in machine learning applications.
6. A method according to claim 1 , comprising buffering the samples received prior to applying the algorithm.
This invention relates to signal processing, specifically methods for handling and processing received samples to improve data accuracy or efficiency. The problem addressed involves the need to manage incoming data samples effectively before applying a processing algorithm, particularly in scenarios where real-time or near-real-time processing is required. The method involves buffering the received samples before applying the algorithm, which allows for better synchronization, error correction, or other preprocessing steps to enhance the algorithm's performance. The buffering step ensures that the samples are stored temporarily, enabling the algorithm to access a consistent set of data points for processing. This approach is particularly useful in applications where data arrives asynchronously or at varying rates, such as in communication systems, sensor networks, or multimedia streaming. By buffering the samples, the method helps mitigate issues like data loss, timing discrepancies, or processing delays, leading to more reliable and accurate results. The algorithm applied after buffering may include filtering, decoding, or other signal processing techniques, depending on the specific application. The invention aims to improve the robustness and efficiency of signal processing systems by ensuring that the algorithm operates on a well-organized and synchronized set of input data.
7. A method according to claim 1 , comprising using at least two samplers to generate the state vectors and, prior to applying the algorithm, receiving the sampled output from each and removing duplicates.
This invention relates to a method for generating state vectors in a computational system, particularly for applications requiring efficient sampling and deduplication. The method addresses the problem of redundant or duplicate data in sampled outputs, which can lead to inefficiencies in processing and analysis. The core technique involves using at least two samplers to independently generate state vectors, which are representations of system states or data points. Before applying a computational algorithm to these vectors, the method receives the sampled outputs from each sampler and removes any duplicate entries. This deduplication step ensures that only unique state vectors are processed, improving computational efficiency and accuracy. The samplers may operate in parallel or sequentially, and the deduplication process can be performed using hash-based techniques, comparison algorithms, or other methods to identify and eliminate duplicates. By reducing redundant data early in the workflow, the method enhances the performance of subsequent algorithms, such as machine learning models, simulations, or data analysis tasks. The invention is particularly useful in fields like data mining, signal processing, and artificial intelligence, where large datasets and high-dimensional state spaces are common.
8. A method according to claim 1 , wherein the second circuitry is implemented as a quantum annealing machine.
A method for optimizing a computational problem using a quantum annealing machine involves solving an optimization problem by configuring a quantum annealing machine to represent the problem as a quantum system. The quantum annealing machine is initialized in a known state and then evolved through a series of quantum states to find a minimum energy configuration, which corresponds to an optimal solution for the problem. The method includes preparing the quantum annealing machine with parameters that encode the problem, such as coupling strengths and biases, and then allowing the system to evolve under quantum dynamics to reach a ground state. The solution is extracted from the final state of the quantum annealing machine, which represents the optimal configuration for the problem. This approach leverages quantum mechanical effects, such as tunneling and superposition, to efficiently explore the solution space and avoid local minima, making it particularly useful for complex optimization problems that are difficult to solve with classical computing methods. The quantum annealing machine is designed to handle problems formulated as quadratic unconstrained binary optimization (QUBO) or Ising model problems, where variables are binary and interactions between them are represented by weights. The method is applicable to various optimization tasks, including machine learning, logistics, and financial modeling, where finding optimal solutions is computationally intensive.
9. A system comprising: a first circuitry configured to store a parameter vector representing an energy function of a network having a plurality of visible units connected using links to a plurality of hidden units, each link being a relationship between two units; and wherein the first circuitry is further configured to receive state vectors from a second circuitry and to apply an algorithm to produce new data, wherein state vectors produced for the network by the second circuitry are sampled from a probability distribution defined by the parameter vector produced by the first circuitry; wherein the sampling further comprises generating candidate state vectors; and wherein the sampling to generate the candidate state vectors is performed independently from an inner loop of the first circuitry used to produce the new data using the state vectors, wherein the inner loop processes the state vectors to iteratively adjust the parameter vector to converge a marginal probability distribution over the visible units towards an example probability distribution; wherein the sampling and producing the new data is performed in parallel and asynchronously.
This invention relates to a machine learning system for training a network model, specifically addressing the computational inefficiency in sampling-based training methods. The system includes a first circuitry that stores a parameter vector representing an energy function of a network with visible and hidden units connected by links, where each link defines a relationship between units. The first circuitry receives state vectors from a second circuitry and applies an algorithm to generate new data. The second circuitry samples state vectors from a probability distribution defined by the parameter vector, producing candidate state vectors independently from the first circuitry's inner loop. The inner loop processes state vectors iteratively to adjust the parameter vector, converging the marginal probability distribution over visible units toward an example probability distribution. The sampling and data production occur in parallel and asynchronously, improving efficiency by decoupling the sampling process from the parameter update loop. This approach reduces computational overhead and accelerates training by allowing concurrent operations. The system is particularly useful in deep learning and probabilistic modeling, where sampling-based methods are commonly used but often suffer from bottlenecks due to sequential processing.
10. A system according to claim 9 , wherein the circuitries are configured to sample and/or produce the new data until a predetermined condition is reached, at which time data is exchanged between the two circuitries, wherein the predetermined condition is based on the converging of the marginal probability distribution over the visible units towards the example probability distribution.
This invention relates to a system for processing data between two interconnected circuitries, addressing the challenge of efficiently exchanging data while ensuring convergence of probability distributions. The system involves two circuitries that sample and generate new data independently until a specific condition is met. This condition is defined by the convergence of the marginal probability distribution over the visible units toward a target example probability distribution. Once convergence is achieved, the system facilitates data exchange between the two circuitries. The circuitries may operate in parallel, each generating or sampling data based on their respective models or configurations. The system ensures that data exchange occurs only when the probability distributions align sufficiently, optimizing computational efficiency and accuracy. This approach is particularly useful in machine learning and statistical modeling, where iterative sampling and convergence are critical for training models or refining data representations. The invention improves upon prior methods by dynamically triggering data exchange based on probabilistic convergence, reducing unnecessary computations and enhancing performance.
11. A system according to claim 10 , wherein the first circuitry is configured to apply a learning algorithm to produce an updated parameter vector and subsequently to send it to the second circuitry for re-sampling when the predetermined condition is reached.
The system operates in the domain of machine learning and data processing, specifically addressing the challenge of efficiently updating and re-sampling parameter vectors in distributed computing environments. The system includes first circuitry and second circuitry, where the first circuitry is responsible for applying a learning algorithm to generate an updated parameter vector. This updated vector is then transmitted to the second circuitry for re-sampling when a predetermined condition is met. The second circuitry is configured to re-sample the updated parameter vector, ensuring that the data remains representative and statistically valid after updates. The system optimizes the learning process by dynamically adjusting parameters and re-sampling data, improving the accuracy and efficiency of machine learning models in distributed systems. The predetermined condition may be based on factors such as convergence thresholds, computational load, or data drift, ensuring timely updates and maintaining model performance. This approach enhances scalability and reliability in large-scale machine learning applications.
12. A system according to claim 11 , wherein the first circuitry is configured to send the updated parameter vector to the second circuitry for re-sampling when a predetermined plural number of iterations of the learning algorithm have been performed.
The invention relates to a machine learning system designed to optimize parameter updates during training. The system addresses the challenge of efficiently refining model parameters by dynamically adjusting the sampling process based on learning progress. The system includes first circuitry that processes an initial parameter vector using a learning algorithm, such as gradient descent, to generate an updated parameter vector. This circuitry is configured to monitor the number of iterations performed by the learning algorithm. When a predetermined number of iterations has been completed, the first circuitry sends the updated parameter vector to second circuitry for re-sampling. The second circuitry then re-samples the updated parameter vector to improve the learning process, ensuring more accurate and efficient parameter updates. This iterative re-sampling mechanism helps enhance model performance by refining the parameter space exploration. The system is particularly useful in scenarios where traditional sampling methods may lead to suboptimal convergence or slow training times. By integrating re-sampling at specific intervals, the system ensures that the learning algorithm remains adaptive and responsive to the evolving parameter landscape.
13. A system according to claim 9 , wherein the first circuitry is configured to take as input the state vectors and the parameter vector, and to generate output data representing a probability distribution over all possible states of the visible units.
This invention relates to a machine learning system, specifically a neural network architecture designed to model complex data distributions. The system addresses the challenge of efficiently representing and sampling from high-dimensional probability distributions, which is critical for applications like generative modeling, data imputation, and uncertainty quantification. The system includes a neural network with visible units representing observable data and hidden units capturing latent factors. A key component is circuitry that processes state vectors (representing hidden unit states) and a parameter vector (containing learned model parameters) to compute a probability distribution over all possible states of the visible units. This allows the system to generate samples or make probabilistic inferences about the visible data given the hidden state. The circuitry operates by combining the state vectors and parameter vector through learned transformations, such as matrix multiplications and nonlinear activations, to produce a distribution that can be sampled. This enables the system to model dependencies between visible units and capture the underlying data structure. The architecture is particularly useful for tasks requiring probabilistic outputs, such as image generation, time-series forecasting, or anomaly detection. The system improves upon prior approaches by providing a flexible and scalable way to compute visible unit distributions, leveraging the hidden state to enhance accuracy and generalization. The parameter vector allows the model to adapt to different datasets and tasks, while the state vectors enable dynamic updates based on observed data. This makes the system suitable for real-world applications where data distributions are complex and evolve ove
14. A system according to claim 9 , wherein the first circuitry is configured to buffer the samples received prior to applying the algorithm.
The system relates to signal processing, specifically for buffering and processing samples using an algorithm. The problem addressed is the need to efficiently handle and process incoming data samples before applying computational algorithms, ensuring accurate and timely results. The system includes circuitry designed to buffer received samples before applying an algorithm to process them. This buffering step allows for temporary storage and synchronization of data, preventing loss or corruption during processing. The algorithm applied to the buffered samples may involve filtering, transformation, or other signal processing techniques to extract meaningful information. The system ensures that the algorithm operates on a stable set of data, improving reliability and performance. The buffering circuitry may include memory components or registers to hold the samples until the algorithm is ready to process them. This approach is particularly useful in real-time applications where data integrity and processing speed are critical. The system may be integrated into larger signal processing frameworks, such as communication systems, sensor networks, or data acquisition platforms, where buffering and algorithmic processing are essential for accurate analysis.
15. A system according to claim 9 , wherein the second circuitry comprises at least two samplers configured to generate the state vectors and, prior to applying the algorithm, to receive the sampled output from each and to remove duplicates.
The system relates to signal processing and data analysis, specifically for reducing redundancy in sampled data to improve computational efficiency. The problem addressed is the presence of duplicate or redundant data points in sampled signals, which can lead to unnecessary processing and inefficiencies in subsequent analysis. The system includes circuitry for sampling and processing signals to generate state vectors, which are mathematical representations of the signal's state at different points in time. The circuitry comprises at least two samplers that independently capture the signal output. Before applying any algorithm to the sampled data, the system removes duplicate entries from the sampled outputs. This ensures that only unique data points are processed, reducing computational overhead and improving the accuracy of the analysis. The samplers operate in parallel or sequentially to capture the signal at different intervals or conditions, generating multiple state vectors. By removing duplicates before further processing, the system avoids redundant calculations and enhances the efficiency of subsequent algorithms, such as machine learning models, signal filtering, or pattern recognition tasks. This approach is particularly useful in applications where real-time processing or high-speed data analysis is required, such as in telecommunications, sensor networks, or industrial automation.
16. A system according to claim 9 , wherein the first and second circuitries are implemented on different hardware circuitries having their own microprocessor or microcontroller.
The invention relates to a system for distributed processing, addressing the challenge of efficiently managing computational tasks across multiple hardware components. The system includes at least two circuitries, each implemented on separate hardware with its own dedicated microprocessor or microcontroller. These circuitries are designed to perform distinct functions, such as data processing, control, or communication, while operating independently. The separation of hardware ensures isolation, reducing interference and improving reliability. Each circuitry can execute tasks autonomously, allowing for parallel processing and enhanced performance. The system may also include interfaces or communication protocols to enable coordination between the circuitries, ensuring seamless integration despite their physical separation. This architecture is particularly useful in applications requiring high reliability, such as industrial automation, embedded systems, or real-time control, where independent operation of components is critical. The use of separate microprocessors or microcontrollers further enhances fault tolerance and scalability, as failures in one circuitry do not necessarily affect the others. The system optimizes resource utilization by distributing workloads across multiple hardware units, improving efficiency and robustness.
17. A system according to claim 9 , wherein the second circuitry is implemented as a quantum annealing machine.
A system for optimizing computational tasks using quantum annealing technology. The system includes a quantum annealing machine configured to solve optimization problems by finding low-energy states in a quantum system. The quantum annealing machine operates by evolving a quantum system from an initial Hamiltonian to a problem Hamiltonian, where the problem Hamiltonian encodes the optimization problem to be solved. The system further includes a classical processing unit that prepares the problem Hamiltonian based on input data and interfaces with the quantum annealing machine to execute the optimization process. The quantum annealing machine is designed to handle complex optimization problems that are computationally intensive for classical computers, leveraging quantum tunneling and superposition to explore solution spaces more efficiently. The system is particularly useful in fields such as machine learning, logistics, and financial modeling, where finding optimal solutions is critical but challenging for classical algorithms. The quantum annealing machine provides a specialized hardware solution that accelerates the optimization process by exploiting quantum mechanical effects, offering potential advantages in speed and solution quality over classical optimization methods.
18. A non-transitory computer-readable storage medium having stored thereon computer-readable code, which, when executed by computing apparatus, causes the computing apparatus to perform a method comprising: storing, by first circuitry, a parameter vector representing an energy function of a network having a plurality of visible units connected using links to a plurality of hidden units, each link being a relationship between two units; and receiving, by the first circuitry, state vectors from second circuitry and applying an algorithm to produce new data, wherein the state vectors produced by the second circuitry are sampled from a probability distribution defined by the parameter vector produced by the first circuitry; and wherein the sampling further comprises generating candidate state vectors; and wherein the sampling to generate the candidate state vectors is performed independently from an inner loop of the first circuitry used to produce the new data using the state vectors, wherein the inner loop processes the state vectors to iteratively adjust the parameter vector to converge a marginal probability distribution over the visible units towards an example probability distribution; wherein the sampling and producing the new data is performed in parallel and asynchronously.
This invention relates to machine learning systems, specifically neural networks with visible and hidden units, addressing the challenge of efficiently sampling state vectors from a probability distribution defined by a parameter vector. The system includes first circuitry that stores a parameter vector representing an energy function of a network with visible and hidden units connected by links. The first circuitry receives state vectors from second circuitry, which samples these vectors from a probability distribution defined by the parameter vector. The sampling process involves generating candidate state vectors independently from an inner loop of the first circuitry, which processes the state vectors to iteratively adjust the parameter vector. This adjustment aims to converge the marginal probability distribution over the visible units toward an example probability distribution. The sampling and data production occur in parallel and asynchronously, improving computational efficiency. The system optimizes the training of neural networks by decoupling the sampling process from the parameter adjustment loop, allowing for faster and more scalable learning.
Unknown
November 24, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.